Rocco Georgi

openai could have simply changed their name many years ago - then they wouldn't be held to FOSS standards which literally no other company is expected to conform to there are fewer demands for open-source windows, open-source MacOS, or even open-sourcing the formula of diet coke or the algorithm behind TikTok (or github itself which was acquired by microsoft for 7.5b), than there are for the weights of GPT-4 and beyond, for the simple reason that everyone else wishes it was theirs and has perpetual sour grape syndrome as a result of being several years too slow to realize scaling laws were the real deal and that conviction and talent actually matter. i say this as someone that heavily uses and loves open-source software, has linux on every desktop and laptop that he owns, and regularly donates to the EFF it has never been remotely feasible to find a way to conjure tens of billions of dollars (or perhaps trillions in the longer run) de novo with the expectation that it will disappear into thin air (unless you are a government) as it is laboriously converted to a few gigabytes of weights which quickly disperse and transmute throughout the internet, then leaving the original creators compensation only in the form of a pile of google scholar citations plus a few kilostars on github revenue models which are able to reliably underpin open source software are in many cases dissimilar from those that function with open source weights (although many still exist!), and it is not a coincidence that every large-scale (1b+) venture-backed ai startup very quickly realized that they would have to be at least partially closed source, or their margins would be driven not just to zero as may happen with many foundational model companies, but even into the negatives, and sometimes within a week of a release which they spent 8 figures of compute and months of work on. this is how you end up with zero revenue and a bankrupt company the moment venture capital markets eye something shinier than you (which may or may not happen soon, but you will eventually go bankrupt if you spend more money than you make) there have been many resulting waves of self-administered therapy to treat this collective sour-grape syndrome which intensifies every time it's realized that companies like openai and anthropic are still winning at model quality. the most recent form of this has been to administer copium directly into our spinal cords by continually fine-tuning and releasing with the goal of beating various benchmarks which a few people made up for fun in their spare time several years ago and which almost no one outside of the evals teams at the labs bothers to even look at, the result of which is many series of twitter posts claiming that a rag-tag team of anime avatars (hey~) and PhDs with treatment-resistant stockholm syndrome have finally beaten the big labs and their faux ideas of 'safety' once and for all, with some new model on a barely-changed-since-literally-2020 architecture which was trained primarily on reddit posts from a decade ago combined with scraped gpt-4 conversations it seems like ai will actually matter in our lifetime, so it makes sense to be wary of the excessive centralization of power which it may help to accrue, and this does seem to be a very important issue. unfortunately, even if you do find a way to produce such artifacts which already cost nine or more figures and to then distribute them freely among the populace, the result will look nothing like what you imagine, because everything will still end up power-law distributed and in more ways than our imaginations are even used to or capable of reasonably modelling to begin with many of the smartest humans alive do not have a particularly enthralling track record of predicting the course AI would take up until this point, so it is reasonable to have low confidence in this area and expect many high-magnitude surprises in the future. that's why i only have conviction in the low-iq plays like buying nvidia, because the rest will be left to whatever molochian psychofauna end up devouring you shortly after you're finally freed from this website consider remaining open to the landscape significantly changing in ways which you do not expect and which may invalidate beliefs which are dear to you. if you have a set of beliefs which you could not possibly imagine invalidating, consider invalidating your meta-logic for choosing your sets of beliefs as well